← Back to Sessions

Towards a Practical Physics-Informed Neural Network Method for End-to-End Full Waveform Inversion

Description: 

Our aim is to explore the viability of end-to-end deep learning (DL) methods for full waveform inversion (FWI) towards practical use for active-source seismic data shot gathers, which may be able to overcome some of the challenges of traditional FWI. We developed a DL-based FWI method that inputs elastic full-waveform shot gathers along a seismic line and outputs a 2-D P-wave velocity model. We employ a physics-informed neural network (PINN) that solves the acoustic wave equation for enhanced generalizability and physics-based inverse solution. This method takes a starting velocity model guess and iterates over solving for the acoustic wave equation via PINN with boundary and initial condition constraints, including a point source location and source function, while fitting the input data for an updated velocity model that honors the physics. We generate a starting velocity model using a data-driven neural network trained on over three thousand velocity model and shot gather pairs. We explore multiple PINN architectures, including a physics-informed (PI)-DeepONet, which is a neural operator that allows for further generalizability and flexibility for enforcing boundary condition constraints (e.g., shot location and source function).

We find that our current methodology requires long training times for each inversion when incorporating a PINN and the solution struggles to converge for the source frequencies of interest. The S-wave energy in our elastic data further complicates our training and appears to confuse the NN and results in poorer solutions. Further research into PINN convergence for the wave equation, ways to limit frequency bias during learning, preprocessing techniques for sseismic data, and training strategies (e.g., domain decomposition, frequency staging), will pave the way for future end-to-end method that can compete with traditional or hybrid FWI methods at the scale and frequencies generally of interest for practical applications.

SNL is managed and operated by NTESS under DOE NNSA contract DE-NA0003525.

Session: Machine Learning for Full Waveform Inversion: From Hybrid to End-to-End Approaches [Poster Session]

Type: Poster

Date: 5/3/2024

Presentation Time: 08:00 AM (local time)

Presenting Author: Jennifer

Student Presenter: No

Invited Presentation: 

Authors